52 1. Preparatory material
a singular value decomposition with singular values σ2 ≥ · · · ≥ σp ≥ 0 and
singular vectors u2,...,up ∈ u1
v2,...,vp ∈ v1
with the stated properties.
By construction we see that σ2,...,σp are less than or equal to σ1. If we
now adjoin σ1,u1,v1 to the other singular values and vectors we obtain the
Exercise 1.3.16. Show that the singular values σ1(A) ≥ · · · ≥ σp(A) ≥ 0
of a p × n matrix A are unique. If we have σ1(A) · · · σp(A) 0, show
that the singular vectors are unique up to rotation by a complex phase.
By construction (and the above uniqueness claim) we see that σi(UAV )
= σi(A) whenever A is a p × n matrix, U is a unitary p × p matrix, and V is
a unitary n × n matrix. Thus the singular spectrum of a matrix is invariant
under left and right unitary transformations.
Exercise 1.3.17. If A is a p × n complex matrix for some 1 ≤ p ≤ n, show
that the augmented matrix
is a p+n×p+n Hermitian matrix whose eigenvalues consist of ±σ1(A),...,
±σp(A), together with n − p copies of the eigenvalue zero. (This generalises
Exercise 2.3.17.) What is the relationship between the singular vectors of A
and the eigenvectors of
Exercise 1.3.18. If A is an n×n Hermitian matrix, show that the singular
values σ1(A),...,σn(A) of A are simply the absolute values |λ1(A)|,...,
|λn(A)| of A, arranged in descending order. Show that the same claim
also holds when A is a normal matrix (that is, when A commutes with
its adjoint). What is the relationship between the singular vectors and
eigenvectors of A?
Remark 1.3.10. When A is not normal, the relationship between eigen-
values and singular values is more subtle. We will discuss this point in later
Exercise 1.3.19. If A is a p × n complex matrix for some 1 ≤ p ≤ n,
together with n − p copies of the eigenvalue zero.
Based on this observation, give an alternate proof of the singular value de-
composition theorem using the spectral theorem for (positive semi-definite)
Exercise 1.3.20. Show that the rank of a p × n matrix is equal to the
number of non-zero singular values.